Prompt Vault admin guide
Prompt Vault lives in Settings > Privacy. It is an organization-level control for raw user prompt retention across supported Runtime Shield prompt events and Claude/Cowork OTLP logs.
Prompt Vault lives in Settings > Privacy. It is an organization-level control for raw user prompt retention across supported Runtime Shield prompt events and Claude/Cowork OTLP logs.
Choosing a mode
Start with Sealed Evidence for most enterprise production rollouts. It preserves forensic access while preventing raw prompts from appearing in normal evidence flows.
Choose Visible Capture when the organization explicitly wants prompt text available to analysts for early pilots, internal security testing, or high-context incident review.
Choose Promptless when policy, contract, or data classification rules say raw prompts must not be stored at all.
Operational impact
| Area | Visible Capture | Sealed Evidence | Promptless |
|---|---|---|---|
| Prompt detection | Full prompt evaluated in memory | Full prompt evaluated in memory | Full prompt evaluated in memory |
| Normal evidence | Raw prompt visible | Metadata plus reveal control | Metadata plus not-retained state |
| OTLP prompt fields | May be retained | Scrubbed from normal OTLP rows | Scrubbed and discarded |
| Reveal | Not needed | Owner/admin with reason | Not available |
| Best audit phrase | Retained and visible | Sealed with audited reveal | Not retained |
Rollout checklist
- Open Settings > Privacy and select the Prompt Vault mode.
- Confirm OTLP privacy settings are also configured on the Privacy tab.
- Generate a benign prompt event from each supported source in scope.
- Open Activity and verify the displayed prompt state matches the selected mode.
- In Sealed Evidence, perform one test reveal with a reason and confirm the log appears.
- In Promptless, confirm there is no Reveal action and no normal API response contains the prompt body.
- Document the chosen mode in the customer rollout notes.
Reveal operations
Only owner and admin users can reveal sealed prompt evidence. A reveal request without a reason is rejected. A successful reveal writes an append-only log before decryption and response.
Developers and legacy Members cannot reveal prompts. Reveal requests must identify the evidence source, evidence ID, prompt field, and reason. Exports remain sealed in this release, so prompt text is only returned through the governed dashboard reveal flow.
Recommended internal reason format:
Investigation CASE-123: validate suspected prompt injection report from workstation ak-mbp-17.
Do not use reveal for casual debugging. Use the visible metadata first: source, session, host, prompt hash, prompt length, detector reason, and surrounding activity often provide enough context.
Migration behavior
Existing organizations remain in Visible Capture unless their legacy prompt logging toggle was already disabled. Those organizations migrate to Promptless.
New enterprise customers should make an explicit Prompt Vault choice during onboarding. For a privacy-forward default, choose Sealed Evidence.
Customer-facing language
Use mode-specific language instead of a blanket statement:
- Visible Capture: "AgentKeeper retains prompt text and authorized dashboard users can view it."
- Sealed Evidence: "AgentKeeper does not show prompt text in normal telemetry. Prompt text is encrypted and requires audited privileged reveal."
- Promptless: "AgentKeeper evaluates supported prompt events in memory and does not retain raw prompt text."
Related detail: showcase, technical coverage, and What AgentKeeper sees.